fix(vllm): preserve cache keys across restarts

[ketor]

Summary

• require a fixed PYTHONHASHSEED together with SHA-256 block hashing because current vLLM seeds the root block hash from os.urandom() when unset
• invalidate process-shared absent rendezvous entries after successful scalar, batch, and SG PUTs
• document the cross-process identity requirement and cover both guards with regression tests

0064 evidence

• before: identical 94k-token prompt after engine restart probed 5,748 candidates with 0 present; first block key changed across restarts
• after PYTHONHASHSEED=0: first block key remained identical and 5,872/5,872 objects were found after restart
• cold SSD retrieval: 35.85 GB useful TP payload in 1.64-1.75 s; nvme0n1 2.11 GB/s + nvme2n1 2.18 GB/s; request 2.82 s; zero RDMA completion errors
• warm retrieval: 35.85 GB useful TP payload in 1.11-1.18 s; request 2.72 s; zero RDMA completion errors

Checks

• python3 -m unittest test_determinism_guard: 6 passed
• client_metrics_test filters for PUT invalidation and REMOVE invalidation: 2 passed
• clean end-to-end restart smoke: 94k-token cached prompt completed in 2.16 s